basin 1.1.0

Numerical optimization in pure Rust, with pluggable linear-algebra backends and WASM support.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
//! The method of RESCUE (BOBYQA, Powell 2009, §5).
//!
//! When rounding errors damage the denominator `σ` of the least-Frobenius-norm
//! `H`-update (eq. 4.9), the interpolation set's geometry has degraded.
//! RESCUE replaces a few interpolation points by fresh ones to restore the
//! conditioning of the interpolation system. It:
//!
//! 1. recentres the origin on `x_opt` ([`QuadraticModel::shift_origin`]);
//! 2. lays down a set of *provisional* interpolation points around `x_opt`
//!    (coordinate steps of length `δ`, clipped to the bounds), represented
//!    implicitly by `ptsid` + `ptsaux`, and rebuilds `H` (`Ξ`/`Υ`/`Ω`) for that
//!    provisional set with all `Ω`-signs `+1`;
//! 3. reinstates as many *original* points as can safely replace a provisional
//!    point (the denominator test of eq. 5.7), updating `H` via [`updateh_rsc`];
//! 4. evaluates `F` at the genuinely-new provisional points that remain and
//!    rebuilds the quadratic model `∇Q` / `Γ` / `γ` by Lagrange corrections.
//!
//! Ported from PRIMA v0.7.2 `fortran/bobyqa/rescue.f90` (the `rescue` and
//! `updateh_rsc` subroutines), translated into basin's [`QuadraticModel`]
//! layout. Two deliberate departures from the literal Fortran, both forced by
//! basin's representation rather than by choice:
//!
//! - **The gradient `gq` stays anchored at `x0`.** PRIMA's `gopt` is `∇Q(x_opt)`
//!   and its final step (rescue.f90:559–561) re-anchors `gopt` when `kopt`
//!   moves. basin stores `∇Q(x0)` and derives `gradient_at_opt()` on demand, so
//!   that re-anchoring is omitted; after the §5 origin shift `x0 = x_opt`, so the
//!   incremental `gq += moderr·Ξ_kpt` updates are exactly right for `∇Q(x0)`.
//! - **No `maxfun` gate.** PRIMA aborts RESCUE when the evaluation budget is
//!   exhausted; basin enforces budgets framework-side (the `Executor`'s
//!   `MaxCostEvals`), so RESCUE evaluates the handful of new points it needs and
//!   lets the executor stop afterward, like [`super::init`] does at start-up.

use crate::core::math::Scalar;
use crate::solver::powell::QuadraticModel;

use super::init::{extra_pair, xinbd};

/// Decode PRIMA's packed `ptsid` value into the 1-based coordinate indices
/// `(ip, iq)`; `0` means "absent" (rescue.f90:358–359). With `ip = ⌊ptsid⌋` and
/// `iq = ⌊(n+1)·ptsid − (n+1)·ip⌋`, the provisional step is
/// `ptsaux(1,ip)·e_ip + ptsaux(1,iq)·e_iq` (both present), or a single
/// `ptsaux(1,ip)·e_ip` / `ptsaux(2,iq)·e_iq` (only `ip` / only `iq`).
fn decode_ptsid<F: Scalar>(p: F, n: usize) -> (usize, usize) {
    let np1 = F::from_usize(n + 1).expect("n+1 representable");
    let ip_f = p.floor();
    let ip = ip_f.to_usize().unwrap_or(0);
    let iq_f = (np1 * p - np1 * ip_f).floor();
    let iq = iq_f.to_usize().unwrap_or(0);
    (ip, iq)
}

/// Plain dot product of two equal-length slices.
fn dot<F: Scalar>(a: &[F], b: &[F]) -> F {
    a.iter().zip(b).map(|(x, y)| *x * *y).sum()
}

/// Run RESCUE (Powell 2009, §5) on `model`, mutating it (interpolation set,
/// `H`-factorization, quadratic model, `kopt`, `x0`) and the shifted bounds
/// `sl`/`su` in place. `lower`/`upper` are the absolute box `[a, b]`, `delta`
/// the current trust-region radius; `eval` evaluates `F` at the genuinely-new
/// provisional points (its first `Err` bubbles), and each `(x_abs, F)` it
/// produces is appended to `evaluated` for the driver to fold into its best.
// The shifted bounds (`sl`/`su`), absolute bounds (`lower`/`upper`), `delta`,
// `eval`, and `evaluated` are distinct disjoint borrows of `BobyqaWork` fields
// at the call site; bundling them into a struct would fight the borrow checker.
#[allow(clippy::too_many_arguments)]
pub(crate) fn rescue<F: Scalar, E>(
    model: &mut QuadraticModel<F>,
    sl: &mut [F],
    su: &mut [F],
    lower: &[F],
    upper: &[F],
    delta: F,
    eval: &mut impl FnMut(&[F]) -> Result<F, E>,
    evaluated: &mut Vec<(Vec<F>, F)>,
) -> Result<(), E> {
    let n = model.n();
    let m = model.m();
    let rank = m - n - 1;
    let zero = F::zero();
    let one = F::one();
    let half = F::from_f64(0.5).expect("0.5 representable");
    let two = F::from_f64(2.0).expect("2.0 representable");
    let np1 = F::from_usize(n + 1).expect("n+1 representable");

    // --- Prologue: shift the origin onto x_opt (rescue.f90:233–244). ---
    let kopt0 = model.kopt();
    let xopt = model.xpt_row(kopt0).to_vec();
    // The shifted bounds move with the new origin (= old x_opt).
    for i in 0..n {
        sl[i] = (sl[i] - xopt[i]).min(zero);
        su[i] = (su[i] - xopt[i]).max(zero);
    }
    // shift_origin recentres x0/xpt/gq/Γ (γ unchanged); its Ξ/Υ work is
    // overwritten below by the provisional H. Equivalent to PRIMA's HQ r2update
    // prologue, and additionally re-anchors gq to the new x0 (= x_opt).
    model.shift_origin();
    // Clamp x0 into the box (PRIMA: xbase = min(max(xl, xbase+xopt), xu)); a
    // no-op in exact arithmetic since x0 + xopt stays feasible.
    for i in 0..n {
        model.x0[i] = lower[i].max(upper[i].min(model.x0[i]));
    }
    // x_opt is now the origin; force its displacement to exactly zero.
    for i in 0..n {
        model.xpt.set(kopt0, i, zero);
    }

    // --- PTSAUX: the two provisional step lengths per coordinate (5.4–5.5;
    // rescue.f90:247–252). ptsaux[j] = [ptsaux(1,j+1), ptsaux(2,j+1)]. ---
    let ptsaux: Vec<[F; 2]> = (0..n)
        .map(|j| {
            let mut a1 = delta.min(su[j]);
            let mut a2 = (-delta).max(sl[j]);
            if a1 + a2 < zero {
                core::mem::swap(&mut a1, &mut a2);
            }
            if a2.abs() < half * a1.abs() {
                a2 = half * a1;
            }
            [a1, a2]
        })
        .collect();

    // --- Provisional H from the coordinate-direction points (rescue.f90:254–276).
    // Identifiers go into `ptsid`; the nonzero Ξ/Υ/Ω entries are written below,
    // with all Ω-signs forced to +1. ---
    let sfrac = half / np1;
    let mut ptsid = vec![zero; m];
    ptsid[0] = sfrac;
    for i in 0..n {
        for j in 0..m {
            model.bmat_xi.set(i, j, zero);
        }
        for j in 0..n {
            model.bmat_ups.set(i, j, zero);
        }
    }
    for i in 0..m {
        for j in 0..rank {
            model.zmat.set(i, j, zero);
        }
    }
    for k in 0..rank {
        model.zsign[k] = one;
    }
    for k in 1..=n {
        let c = k - 1; // 0-based coordinate
        let kf = F::from_usize(k).expect("k representable");
        ptsid[k] = kf + sfrac;
        if k < m - n {
            ptsid[k + n] = kf / np1 + sfrac;
            let temp = one / (ptsaux[c][0] - ptsaux[c][1]);
            model.bmat_xi.set(c, k, -temp + one / ptsaux[c][0]);
            model.bmat_xi.set(c, k + n, temp + one / ptsaux[c][1]);
            let b0 = -model.bmat_xi.get(c, k) - model.bmat_xi.get(c, k + n);
            model.bmat_xi.set(c, 0, b0);
            let z0 = two.sqrt() / (ptsaux[c][0] * ptsaux[c][1]).abs();
            model.zmat.set(0, c, z0);
            model.zmat.set(k, c, z0 * ptsaux[c][1] * temp);
            model.zmat.set(k + n, c, -z0 * ptsaux[c][0] * temp);
        } else {
            // Unreachable for m >= 2n+1 (basin's only supported regime); ported
            // from rescue.f90:271–275 for completeness.
            model.bmat_xi.set(c, 0, -one / ptsaux[c][0]);
            model.bmat_xi.set(c, k, one / ptsaux[c][0]);
            model
                .bmat_ups
                .set(c, c, -half * ptsaux[c][0] * ptsaux[c][0]);
        }
    }

    // --- Provisional H from the two-coordinate points (rescue.f90:278–287),
    // reusing `extra_pair` as PRIMA's `setij`. ---
    for t in 0..(m - 2 * n - 1) {
        let slot = 2 * n + 1 + t; // 0-based provisional point
        let (p, q) = extra_pair(t, n); // 0-based coords; PRIMA ip=p+1, iq=q+1
        let ipf = F::from_usize(p + 1).expect("ip representable");
        let iqf = F::from_usize(q + 1).expect("iq representable");
        ptsid[slot] = ipf + iqf / np1 + sfrac;
        let temp = one / (ptsaux[p][0] * ptsaux[q][0]);
        let col = slot - n - 1; // 0-based Ω column for this extra point
        model.zmat.set(0, col, temp);
        model.zmat.set(slot, col, temp);
        model.zmat.set(p + 1, col, -temp);
        model.zmat.set(q + 1, col, -temp);
    }

    // --- Reinstate x_opt at provisional slot kopt by exchanging slots 1 and
    // kopt; mark slot kopt as taken (rescue.f90:289–297). ---
    if kopt0 != 0 {
        for i in 0..n {
            let a = model.bmat_xi.get(i, 0);
            let b = model.bmat_xi.get(i, kopt0);
            model.bmat_xi.set(i, 0, b);
            model.bmat_xi.set(i, kopt0, a);
        }
        for j in 0..rank {
            let a = model.zmat.get(0, j);
            let b = model.zmat.get(kopt0, j);
            model.zmat.set(0, j, b);
            model.zmat.set(kopt0, j, a);
        }
    }
    ptsid[0] = ptsid[kopt0];
    ptsid[kopt0] = zero;

    // --- Scores: distance of each original point from x_opt (rescue.f90:299–312).
    // The paper uses the distance (not its square); SCORE(KOPT)=0 skips x_opt. ---
    let mut score: Vec<F> = (0..m)
        .map(|k| model.xpt_row(k).iter().fold(zero, |a, &v| a + v * v).sqrt())
        .collect();
    score[kopt0] = zero;
    let scoreinc = score.iter().fold(zero, |a, &v| a.max(v));
    let mut nprov = m - 1;

    // --- The reinstatement loop (rescue.f90:326–447). Bounded by m² iterations;
    // retains at least one provisional point. ---
    let c005 = F::from_f64(5e-2).expect("0.05 representable");
    let max_iter = m * m;
    for _ in 0..max_iter {
        if score.iter().all(|&s| s <= zero) || nprov <= 1 {
            break;
        }

        // KORIG: the closest original point not yet processed (smallest positive
        // score).
        let mut korig = 0;
        let mut best = F::infinity();
        for (k, &s) in score.iter().enumerate() {
            if s > zero && s < best {
                best = s;
                korig = k;
            }
        }

        // WMV = w(XNEW) − w(XOPT) for XNEW = XPT(:,KORIG), XOPT = 0
        // (rescue.f90:352–373). Provisional points are read through ptsid/ptsaux.
        let xkorig = model.xpt_row(korig).to_vec();
        let mut wmv = vec![zero; m + n];
        for (k, w) in wmv.iter_mut().enumerate().take(m) {
            let val = if k == kopt0 {
                zero
            } else if ptsid[k] <= zero {
                dot(&xkorig, model.xpt_row(k))
            } else {
                let (ip, iq) = decode_ptsid(ptsid[k], n);
                if ip > 0 && iq > 0 {
                    xkorig[ip - 1] * ptsaux[ip - 1][0] + xkorig[iq - 1] * ptsaux[iq - 1][0]
                } else if ip > 0 {
                    xkorig[ip - 1] * ptsaux[ip - 1][0]
                } else if iq > 0 {
                    xkorig[iq - 1] * ptsaux[iq - 1][1]
                } else {
                    zero
                }
            };
            *w = half * val * val;
        }
        wmv[m..(m + n)].copy_from_slice(&xkorig);

        // VLAG = H·WMV + e_KOPT (rescue.f90:375–388); β (4.10) of BOBYQA paper.
        let (mut vlag_lambda, vlag_g) = model.apply_h(&wmv[..m], &wmv[m..]);
        vlag_lambda[kopt0] = vlag_lambda[kopt0] + one;
        // bsum = wmv(1:n) · (Ξ·wmv_λ + vlag_g), with vlag_g = Ξ·wmv_λ + Υ·wmv_g.
        let mut xi_wmv = vec![zero; n];
        for (i, x) in xi_wmv.iter_mut().enumerate() {
            let mut acc = zero;
            for j in 0..m {
                acc = acc + model.bmat_xi.get(i, j) * wmv[j];
            }
            *x = acc;
        }
        // wᵀHw = ‖Z'wmv_λ‖² + wmv_g·(2·Ξ·wmv_λ + Υ·wmv_g), the cross/Υ terms
        // being `bsum`. N.B. basin contracts against the g-part `wmv_g`
        // (`wmv[m..]`), the exact form per BOBYQA eq. 4.12 and the PRIMA comment's
        // own derivation; PRIMA's rescue.f90:384 contracts against `wmv(1:n)`
        // (the first n λ-entries), which only meets its loose internal 1e-2
        // H-error check. The g-part form makes the rebuilt H equal inv(W) to
        // working precision (asserted by `assert_h_matches_inverse` in tests).
        let mut bsum = zero;
        for i in 0..n {
            bsum = bsum + wmv[m + i] * (xi_wmv[i] + vlag_g[i]);
        }
        let mut zsum = zero;
        for j in 0..rank {
            let mut zw = zero;
            for i in 0..m {
                zw = zw + model.zmat.get(i, j) * wmv[i];
            }
            zsum = zsum + zw * zw;
        }
        let norm_sq = xkorig.iter().fold(zero, |a, &v| a + v * v);
        let beta = half * norm_sq * norm_sq - zsum - bsum;

        // DEN(k) = σ for replacing provisional point k (rescue.f90:392–394).
        let mut den = vec![zero; m];
        for (k, d) in den.iter_mut().enumerate() {
            if ptsid[k] > zero {
                let mut hdiag = zero;
                for j in 0..rank {
                    let z = model.zmat.get(k, j);
                    hdiag = hdiag + z * z;
                }
                *d = hdiag * beta + vlag_lambda[k] * vlag_lambda[k];
            }
        }

        // Denominator test (rescue.f90:414–419). If no provisional point can be
        // safely replaced, lower KORIG's priority and retry later.
        let sum_abs = vlag_lambda
            .iter()
            .chain(vlag_g.iter())
            .fold(zero, |a, &v| a + v.abs());
        let maxsq = vlag_lambda.iter().fold(zero, |a, &v| a.max(v * v));
        let any_good = den.iter().any(|&d| d > c005 * maxsq);
        if !(sum_abs.is_finite() && any_good) {
            score[korig] = -score[korig] - scoreinc;
            continue;
        }

        // KPROV: the provisional point with the largest (finite) denominator.
        let mut kprov = 0;
        let mut bestden = F::neg_infinity();
        for (k, &d) in den.iter().enumerate() {
            if !d.is_nan() && d > bestden {
                bestden = d;
                kprov = k;
            }
        }

        // Exchange provisional points KPROV and KORIG, then replace the (now
        // KORIG-th) provisional point by the KORIG-th original point
        // (rescue.f90:423–446).
        if kprov != korig {
            for i in 0..n {
                let a = model.bmat_xi.get(i, kprov);
                let b = model.bmat_xi.get(i, korig);
                model.bmat_xi.set(i, kprov, b);
                model.bmat_xi.set(i, korig, a);
            }
            for j in 0..rank {
                let a = model.zmat.get(kprov, j);
                let b = model.zmat.get(korig, j);
                model.zmat.set(kprov, j, b);
                model.zmat.set(korig, j, a);
            }
            vlag_lambda.swap(kprov, korig);
        }
        ptsid[kprov] = ptsid[korig];
        ptsid[korig] = zero;
        score[korig] = zero;
        for s in score.iter_mut() {
            *s = s.abs();
        }

        let mut vlag_full = vlag_lambda;
        vlag_full.extend_from_slice(&vlag_g);
        updateh_rsc(model, korig, beta, &mut vlag_full);
        nprov -= 1;
    }

    // --- Re-evaluate the remaining provisional (genuinely new) points and
    // rebuild the quadratic model (rescue.f90:455–561). ---
    let kbase = model.kopt();
    let fbase = model.fval[kbase];
    if nprov > 0 {
        for kpt in 0..m {
            if ptsid[kpt] <= zero {
                continue;
            }

            // Absorb γ_kpt into the explicit Hessian, then zero it (Phase B).
            let g_kpt = model.gamma[kpt];
            if g_kpt != zero {
                let row = model.xpt_row(kpt).to_vec();
                for i in 0..n {
                    for j in 0..n {
                        let add = g_kpt * row[i] * row[j];
                        model
                            .gamma_explicit
                            .set(i, j, model.gamma_explicit.get(i, j) + add);
                    }
                }
            }
            model.gamma[kpt] = zero;

            // Materialize the new provisional point (≤ 2 nonzeros).
            let (ip, iq) = decode_ptsid(ptsid[kpt], n);
            for i in 0..n {
                model.xpt.set(kpt, i, zero);
            }
            let mut xp = zero;
            let mut xq = zero;
            if ip > 0 && iq > 0 {
                xp = ptsaux[ip - 1][0];
                model.xpt.set(kpt, ip - 1, xp);
                xq = ptsaux[iq - 1][0];
                model.xpt.set(kpt, iq - 1, xq);
            } else if ip > 0 {
                xp = ptsaux[ip - 1][0];
                model.xpt.set(kpt, ip - 1, xp);
            } else if iq > 0 {
                xq = ptsaux[iq - 1][1];
                model.xpt.set(kpt, iq - 1, xq);
            }

            // Evaluate F at the new point.
            let disp = model.xpt_row(kpt).to_vec();
            let xabs = xinbd(&model.x0, &disp, lower, upper, sl, su);
            let f = eval(&xabs)?;
            evaluated.push((xabs, f));
            model.fval[kpt] = f;
            if f < model.fval[model.kopt] {
                model.kopt = kpt;
            }

            // Model value VQUAD at the new point (its ≤ 2 nonzeros) + implicit part.
            let mut vquad = fbase;
            if ip > 0 && iq > 0 {
                let (p, qq) = (ip - 1, iq - 1);
                vquad = vquad + xp * (model.gq[p] + half * xp * model.gamma_explicit.get(p, p));
                vquad = vquad + xq * (model.gq[qq] + half * xq * model.gamma_explicit.get(qq, qq));
                vquad = vquad + xp * xq * model.gamma_explicit.get(p, qq);
            } else if ip > 0 {
                let p = ip - 1;
                vquad = vquad + xp * (model.gq[p] + half * xp * model.gamma_explicit.get(p, p));
            } else if iq > 0 {
                let qq = iq - 1;
                vquad = vquad + xq * (model.gq[qq] + half * xq * model.gamma_explicit.get(qq, qq));
            }
            let xnew = model.xpt_row(kpt).to_vec();
            let mut implicit = zero;
            for k in 0..m {
                let xxpt = dot(&xnew, model.xpt_row(k));
                implicit = implicit + model.gamma[k] * xxpt * xxpt;
            }
            vquad = vquad + half * implicit;

            // Model update (rescue.f90:532–554).
            let moderr = f - vquad;
            for i in 0..n {
                model.gq[i] = model.gq[i] + moderr * model.bmat_xi.get(i, kpt);
            }
            let mut pqinc = vec![zero; m];
            for (k, pk) in pqinc.iter_mut().enumerate() {
                let mut om = zero;
                for j in 0..rank {
                    om = om + model.zmat.get(k, j) * model.zmat.get(kpt, j);
                }
                *pk = moderr * om;
            }
            for k in 0..m {
                if ptsid[k] <= zero {
                    model.gamma[k] = model.gamma[k] + pqinc[k];
                }
            }
            for k in 0..m {
                if ptsid[k] <= zero {
                    continue;
                }
                let (ipk, iqk) = decode_ptsid(ptsid[k], n);
                if ipk > 0 && iqk > 0 {
                    let (p, qq) = (ipk - 1, iqk - 1);
                    let (ap, aq) = (ptsaux[p][0], ptsaux[qq][0]);
                    let dpp = model.gamma_explicit.get(p, p) + pqinc[k] * ap * ap;
                    model.gamma_explicit.set(p, p, dpp);
                    let dqq = model.gamma_explicit.get(qq, qq) + pqinc[k] * aq * aq;
                    model.gamma_explicit.set(qq, qq, dqq);
                    let dpq = model.gamma_explicit.get(p, qq) + pqinc[k] * ap * aq;
                    model.gamma_explicit.set(p, qq, dpq);
                    model.gamma_explicit.set(qq, p, dpq);
                } else if ipk > 0 {
                    let p = ipk - 1;
                    let ap = ptsaux[p][0];
                    let dpp = model.gamma_explicit.get(p, p) + pqinc[k] * ap * ap;
                    model.gamma_explicit.set(p, p, dpp);
                } else if iqk > 0 {
                    let qq = iqk - 1;
                    let aq = ptsaux[qq][1];
                    let dqq = model.gamma_explicit.get(qq, qq) + pqinc[k] * aq * aq;
                    model.gamma_explicit.set(qq, qq, dqq);
                }
            }
            ptsid[kpt] = zero;
        }
    }
    // Phase D (rescue.f90:559–561) is intentionally omitted: basin's gq is
    // anchored at x0, and after the prologue shift x0 = x_opt, so the
    // incremental gq updates above already give ∇Q(x0). gradient_at_opt()
    // recovers ∇Q(x_kopt) on demand for any moved kopt.

    Ok(())
}

/// Update `Ξ`/`Υ`/`Ω` to reinstate the `knew`-th original point in place of the
/// `knew`-th provisional point (BOBYQA eq. 4.9/4.14; PRIMA `updateh_rsc`,
/// rescue.f90:612–758). Operates on the all-`+1`-sign provisional `Ω`
/// (`zsign` untouched); `fval` and the model are not touched here. `vlag` is the
/// (length `m+n`) `H·w` in `[λ; g]` layout; it is mutated locally. On
/// `DAMAGING_ROUNDING` (non-finite `vlag`/`β` or `denom ≤ 0`) the update is
/// silently skipped to protect `H` — matching PRIMA, which calls this without
/// capturing its optional `info` here (the entry gate at the call site already
/// rejects bad denominators, so the skip is not expected to trigger).
fn updateh_rsc<F: Scalar>(model: &mut QuadraticModel<F>, knew: usize, beta: F, vlag: &mut [F]) {
    let n = model.n();
    let m = model.m();
    let rank = m - n - 1;
    let zero = F::zero();
    let one = F::one();

    let tau = vlag[knew];
    let mut denom = zero;
    for j in 0..rank {
        let z = model.zmat.get(knew, j);
        denom = denom + z * z;
    }
    denom = denom * beta + tau * tau;

    // Rounding can leave VLAG/β non-finite or DENOM ≤ 0; skip rather than wreck H.
    let sum_abs = vlag.iter().fold(zero, |a, &v| a + v.abs());
    if !((sum_abs + beta.abs()).is_finite() && denom > zero) {
        return;
    }

    // After this, VLAG = H·w − e_KNEW.
    vlag[knew] = vlag[knew] - one;

    // Collapse the KNEW-th row of Z onto column 0 with Givens rotations.
    let mut zmax = zero;
    for i in 0..m {
        for j in 0..rank {
            zmax = zmax.max(model.zmat.get(i, j).abs());
        }
    }
    let ztest = F::from_f64(1e-20).expect("1e-20 representable") * zmax;
    for j in 1..rank {
        if model.zmat.get(knew, j).abs() > ztest {
            let a = model.zmat.get(knew, 0);
            let b = model.zmat.get(knew, j);
            let d = (a * a + b * b).sqrt();
            let cos = a / d;
            let sin = b / d;
            model.rotate_zmat_cols(0, j, cos, sin);
        }
        model.zmat.set(knew, j, zero);
    }

    // HCOL = the KNEW-th column of the un-updated H (minus its suppressed entry).
    let zk0 = model.zmat.get(knew, 0);
    let mut hcol = vec![zero; m + n];
    for (i, h) in hcol.iter_mut().enumerate().take(m) {
        *h = zk0 * model.zmat.get(i, 0);
    }
    for r in 0..n {
        hcol[m + r] = model.bmat_xi.get(r, knew);
    }

    // Complete ZMAT (BOBYQA eq. 4.14) and BMAT (last N rows of eq. 4.9).
    let sqrtdn = denom.sqrt();
    for i in 0..m {
        let z = model.zmat.get(i, 0);
        model
            .zmat
            .set(i, 0, (tau / sqrtdn) * z - (zk0 / sqrtdn) * vlag[i]);
    }
    let alpha = hcol[knew];
    let mut v1 = vec![zero; n];
    let mut v2 = vec![zero; n];
    for r in 0..n {
        let vg = vlag[m + r];
        let hg = hcol[m + r];
        v1[r] = (alpha * vg - tau * hg) / denom;
        v2[r] = (-beta * hg - tau * vg) / denom;
    }
    for r in 0..n {
        for j in 0..m {
            let add = v1[r] * vlag[j] + v2[r] * hcol[j];
            model.bmat_xi.set(r, j, model.bmat_xi.get(r, j) + add);
        }
        for j in 0..n {
            let add = v1[r] * vlag[m + j] + v2[r] * hcol[m + j];
            model.bmat_ups.set(r, j, model.bmat_ups.get(r, j) + add);
        }
    }
    // The rank-2 update need not keep Υ exactly symmetric; restore it.
    let half = F::from_f64(0.5).expect("0.5 representable");
    for r in 0..n {
        for c in (r + 1)..n {
            let avg = half * (model.bmat_ups.get(r, c) + model.bmat_ups.get(c, r));
            model.bmat_ups.set(r, c, avg);
            model.bmat_ups.set(c, r, avg);
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::solver::powell::kkt::assert_h_matches_inverse;

    /// `decode_ptsid` must invert the `ptsid` packing for the origin, the
    /// single-coordinate points, and the two-coordinate points, at several `n`.
    #[test]
    fn ptsid_round_trips() {
        for n in [1usize, 2, 3, 5] {
            let np1 = (n + 1) as f64;
            let sfrac = 0.5 / np1;
            // Origin: ip = iq = 0.
            assert_eq!(decode_ptsid(sfrac, n), (0, 0));
            // Single-coordinate "+" point along e_k: ip = k, iq = 0.
            for k in 1..=n {
                let pid = k as f64 + sfrac;
                assert_eq!(decode_ptsid(pid, n), (k, 0), "n={n} ip-only k={k}");
            }
            // Single-coordinate "−" point along e_k: ip = 0, iq = k.
            for k in 1..=n {
                let pid = k as f64 / np1 + sfrac;
                assert_eq!(decode_ptsid(pid, n), (0, k), "n={n} iq-only k={k}");
            }
            // Two-coordinate point (ip, iq), both in 1..=n.
            for ip in 1..=n {
                for iq in 1..=n {
                    let pid = ip as f64 + iq as f64 / np1 + sfrac;
                    assert_eq!(decode_ptsid(pid, n), (ip, iq), "n={n} pair ({ip},{iq})");
                }
            }
        }
    }

    /// End-to-end: after RESCUE rebuilds the interpolation set and model, the
    /// stored `H` must equal `inv(W)` for the new geometry (KKT identity, with
    /// `Ω`-signs all +1), and the model must still interpolate `F` at every
    /// point. This single check validates the provisional-H construction, the
    /// reinstatement loop, `updateh_rsc`, and the model rebuild together.
    fn rescue_rebuilds_and_interpolates(n: usize, m: usize, rho: f64, f: impl Fn(&[f64]) -> f64) {
        let lower = vec![-5.0; n];
        let upper = vec![5.0; n];
        let x0 = vec![0.5; n];
        let mut init = QuadraticModel::try_initialize_bounded::<core::convert::Infallible>(
            x0,
            &lower,
            &upper,
            rho,
            m,
            &mut |x| Ok(f(x)),
        )
        .expect("infallible objective");

        let mut evaluated = Vec::new();
        rescue::<f64, core::convert::Infallible>(
            &mut init.model,
            &mut init.sl,
            &mut init.su,
            &lower,
            &upper,
            rho,
            &mut |x| Ok(f(x)),
            &mut evaluated,
        )
        .expect("infallible objective");

        // KKT identity for the rebuilt geometry.
        assert_h_matches_inverse(&init.model, 1e-8);

        // Interpolation: Q(x_j) − Q(x_opt) == F(x_j) − F(x_opt) for all j.
        let model = &init.model;
        let kopt = model.kopt();
        let q_opt = model.eval_change(model.xpt_row(kopt));
        let f_opt = model.fval(kopt);
        for j in 0..model.m() {
            let qj = model.eval_change(model.xpt_row(j));
            let lhs = qj - q_opt;
            let rhs = model.fval(j) - f_opt;
            assert!(
                (lhs - rhs).abs() < 1e-7,
                "interp j={j}: {lhs} vs {rhs} (Δ={:e})",
                (lhs - rhs).abs()
            );
        }
        // KOPT is the best point.
        for j in 0..model.m() {
            assert!(model.fval(j) >= f_opt - 1e-12, "fval[{j}] below fopt");
        }
    }

    #[test]
    fn rescue_quadratic_2d() {
        // A convex quadratic with a cross term.
        let f = |x: &[f64]| {
            2.0 * x[0] * x[0] + 1.5 * x[0] * x[1] + 3.0 * x[1] * x[1] + x[0] - 2.0 * x[1] + 0.7
        };
        rescue_rebuilds_and_interpolates(2, 5, 0.3, f);
    }

    #[test]
    fn rescue_quadratic_2d_full_npt() {
        let f = |x: &[f64]| (x[0] - 0.3).powi(2) + 2.0 * (x[1] + 0.4).powi(2) + 0.5 * x[0] * x[1];
        rescue_rebuilds_and_interpolates(2, 6, 0.25, f);
    }

    #[test]
    fn rescue_nonquadratic_3d() {
        // Non-quadratic: the model becomes approximate, but the H-algebra and the
        // interpolation conditions are exact regardless of F.
        let f = |x: &[f64]| {
            (1.0 - x[0]).powi(2) + 100.0 * (x[1] - x[0] * x[0]).powi(2) + (x[2] - 0.5).powi(4)
        };
        rescue_rebuilds_and_interpolates(3, 7, 0.4, f);
    }
}